Anthropic published a guest post on Friday, and the authors are from Germany: Jakob Macke, Daniel Gedon and Stefan Wahl work at the ELLIS Institute Tübingen and the University of Tübingen. They hold Anthropic AI for Science grants, and they write about what AI can and cannot do in scientific modeling. It pairs well with Thursday’s analysis of 207,000 research sessions, except that here the researchers speak for themselves.
The piece opens in 1951. Cambridge University’s computer was out of action for six months, so Andrew Huxley had to solve the equations he and Alan Hodgkin had built to describe a squid nerve cell firing on a hand-cranked mechanical calculator. It took him over three weeks.
The fourth power was a guess
One part of the Hodgkin-Huxley model was an educated guess. To make the curve fit their measurements, the two had to raise one variable to the fourth power. They did not know why. Their hypothesis was that several charged particles in the membrane had to move into place before a gate would open. Decades later, long after their 1963 Nobel Prize, another team found the structure of a related potassium channel: four identical pieces arranged around a central pore.
That is the authors’ point. An AI system today would find such an equation in an afternoon, possibly with a closer fit. But what mattered was not knowing that the fourth power worked. It was knowing that a simpler version did not. They quote Richard Hamming from 1962: the purpose of computing is insight, not numbers.
Three steps, and the machine is in all of them
Modeling breaks into building, constraining and interpreting. Constraining is where machines have come furthest. Bayesian inference returns a whole distribution rather than one set of numbers that fits — it is just not computable for complex models. The way around it is simulation-based inference, or SBI: pick random parameters, run the model, record the output, roughly 100,000 times for something like Hodgkin-Huxley and millions of times for anything larger. Then train a network to return the parameters that could have produced a given output.
The simulations cost weeks, the inference afterwards costs seconds. That gap changes what is possible: when two neutron stars collide, SBI can work out roughly where in the sky it happened while the signal is still detectable, in time to point telescopes at it. The same method is running in cosmology, structural biology and geophysics.
For building the model itself, LLM agents now write the candidates, revise them and try again. Here comes the line that separates this post from a brochure: run the agent twice on the same data and you get two different, equally good models. An agent that hands back only one is hiding the rest.
The catch is in their own text
The authors spell it out: these systems have mostly been tested on rediscovering models scientists already know. Evidence that they work on problems with no known answer is sparse.
For a post on a model vendor’s website that is unusually candid, and it is the most useful part. The lesson for anyone analyzing data with Claude is not about model quality but about handling: a single result is not a finding. It is the spread of explanations that fit the same data which tells you where the evidence is thin and which experiment is missing next.